Fix false positive tracer leaks in flax library.

We'd like flax itself never to accidentally trigger the jax leak detector when enabled,
but over the years a few longer-lived temporaries were introduced to the flax codebase
that do trigger it.  We remedy those cases:

- in flax.core remove a redundant use of nonlocal scopes in pack implementation.
- for nn.jit have the hashing helper datastructure maintain a weakref to module.
- explicitly exempt a context capture used in module summary tabulation routine.

PiperOrigin-RevId: 679881642

flax/core/lift.py

@@ -127,7 +127,7 @@ def _partial_pack(
     out_variable_filters: Sequence[CollectionFilter],
     rng_filters: Sequence[PRNGSequenceFilter],
     name=None,
-) -> tuple[Callable[..., Any], Callable[..., Any], Any, Any, Callable[..., Any], Callable[..., Any]]:
+) -> tuple[Callable[..., Any], Callable[..., Any], Any, Any, Callable[..., Any]]:
   """Pack variables and rngs for functional transformations.
 
   The _partial_pack function is the building block for all other lifted transformations.
@@ -175,16 +175,11 @@ def _partial_pack(
     inner_rng_counters.append(rng_counters)
   rng_groups_xs_t = _transpose(rng_groups_xs)
 
-  inner_scopes: list[Scope] = []
-
   def scope_fn(
       variable_groups_xs_t,
       rng_groups_xs_t,
       mutable_filter: CollectionFilter = True,
   ):
-    nonlocal inner_scopes
-    for inner_scope in inner_scopes:
-      inner_scope.invalidate()
     inner_scopes = []
     mutable: Filter = False
     for out_filter in out_variable_filters:
@@ -260,10 +255,6 @@ def repack_fn(inner_scope_tree):
 
     return _transpose(out_variable_groups_xs)
 
-  def invalidate_scopes_fn():
-    for inner_scope in inner_scopes:
-      inner_scope.invalidate()
-
   def publish_results_fn(out_variable_groups_xs_t):
     out_variable_groups_xs = _transpose(out_variable_groups_xs_t)
     for scope, out_variable_groups, rng_counters in zip(
@@ -278,14 +269,14 @@ def publish_results_fn(out_variable_groups_xs_t):
             scope.put_variable(col_name, var_name, value)
 
   return (
-        scope_fn,
-        repack_fn,
-        variable_groups_xs_t,
-        rng_groups_xs_t,
-        publish_results_fn,
-        invalidate_scopes_fn,
+      scope_fn,
+      repack_fn,
+      variable_groups_xs_t,
+      rng_groups_xs_t,
+      publish_results_fn,
     )
 
+
 def pack(
     fn: Callable[..., Any],
     in_variable_filters: Sequence[CollectionFilter],
@@ -322,24 +313,20 @@ def wrapper(scope_tree: Scope, *args, **kwargs):
         variable_groups_xs_t,
         rng_groups_xs_t,
         publish_results_fn,
-        invalidate_scopes_fn,
     ) = _partial_pack(scope_tree, in_variable_filters, out_variable_filters, rng_filters, name)
-    try:
-      if enable_kwargs:
-        y, out_variable_groups_xs_t = fn(
-            scope_fn,
-            repack_fn,
-            variable_groups_xs_t,
-            rng_groups_xs_t,
-            *args,
-            **kwargs,
-        )
-      else:
-        y, out_variable_groups_xs_t = fn(
-            scope_fn, repack_fn, variable_groups_xs_t, rng_groups_xs_t, *args
-        )
-    finally:
-      invalidate_scopes_fn()
+    if enable_kwargs:
+      y, out_variable_groups_xs_t = fn(
+          scope_fn,
+          repack_fn,
+          variable_groups_xs_t,
+          rng_groups_xs_t,
+          *args,
+          **kwargs,
+      )
+    else:
+      y, out_variable_groups_xs_t = fn(
+          scope_fn, repack_fn, variable_groups_xs_t, rng_groups_xs_t, *args
+      )
     publish_results_fn(out_variable_groups_xs_t)
     return y
 

flax/linen/summary.py

@@ -437,8 +437,9 @@ def _get_table_fn(*args, **kwargs):
 
       def _get_variables():
         return module.init(*args, **kwargs)
-
-      variables = jax.eval_shape(_get_variables)
+      # TODO(cgarciae): is it possible to avoid leaking tracers for summaries?
+      with jax.check_tracer_leaks(False):
+        variables = jax.eval_shape(_get_variables)
       calls = module_lib._context.call_info_stack[-1].calls
       calls.sort(key=lambda c: c.index)
 

flax/linen/transforms.py

@@ -31,6 +31,7 @@
   TypeVar,
   Union,
 )
+import weakref
 from collections.abc import Callable, Iterable, Mapping, Sequence
 
 from flax import core
@@ -440,21 +441,25 @@ class _HashableProxy:
   function should be retraced or not
   """
 
-  module: Module
+  module_ref: weakref.ref
   hash_key: int
 
   @classmethod
   def from_module(cls, module: Module) -> '_HashableProxy':
     fingerprint = _module_fingerprint(module)
     hash_key = hash(fingerprint)
-    return cls(module, hash_key)
+    return cls(weakref.ref(module), hash_key)
 
   def __hash__(self):
     return self.hash_key
 
   def __eq__(self, other):
     return isinstance(other, _HashableProxy) and self.hash_key == other.hash_key
 
+  @property
+  def module(self):
+    return self.module_ref()
+
 
 def _module_fingerprint(module: Module) -> tuple[type[Any], Any]:
   return _fingerprint_recursive(module, (), {})

tests/linen/linen_transforms_test.py

@@ -2627,6 +2627,25 @@ def __call__(self, x):
     y = Top().apply(vs, jnp.ones((2, 5)))
     assert vs['aux']['vfoo']['v'].value.shape == ()
 
+  def test_vjp_tracer_leak(self):
+    class LearnScale(nn.Module):
+      @nn.compact
+      def __call__(self, x):
+        p = self.param('scale', nn.initializers.zeros, ())
+        return p * x
+    class Foo(nn.Module):
+      @nn.compact
+      def __call__(self, x):
+        y, bwd = nn.vjp(lambda mdl, x: mdl(x), LearnScale(), x)
+        params_grad, x_grad = bwd(jnp.ones(y.shape))
+        return y, params_grad, x_grad
+    key = jax.random.PRNGKey(0)
+    x = jnp.ones((2, 3))
+    foo = Foo()
+    with jax.check_tracer_leaks():
+      params = foo.init(key, x)
+      foo.apply(params, x)
+
 
 if __name__ == '__main__':
   absltest.main()